An AI knowledge base for property management is the controlled set of answers your AI is allowed to give residents. Without one, a chatbot guesses at your bylaws and your gym hours. URBI takes the other path: Arthur answers from the building's own records and a knowledge base management curates.
What happens when an AI does not know the answer?
It usually answers anyway. A language model produces the most plausible next sentence. It does not stop when the facts run out. OpenAI's own research says models are "encouraged to guess rather than say 'I don't know'" because scoring rewards a confident attempt over an admission.
Now put that in front of your residents. A general chatbot bolted onto a building website has no idea:
- What your bylaws say about short term rentals, pets, or balcony barbecues
- When the gym closes, whether the party room needs a deposit, or which elevator is booked Saturday
- Which vendor handles the boilers, and whether last week's outage is already ticketed
- Whether the person asking even lives in the building
It will still answer all of those. Fluent and wrong is worse than silence, because a resident acts on it. They skip a deposit, book a room that is not free, or quote a rule your board never passed. A manager spends a morning unwinding it. We made the same case in using ChatGPT for leases and bylaws.
So the useful question is not whether the AI is smart. It is what the AI does at the edge of what it knows.
Who decides what the AI is allowed to tell your residents?
With URBI, management does, and that is the design of Arthur's handoff. Arthur is the resident facing AI on phone, SMS, email, and in app chat, in ten languages. Asked something it has no grounded answer for, Arthur does not improvise. It escalates to the property manager with a summary of the question.
The loop closes like this:
- The manager answers once, in their own words.
- Arthur adapts that guidance into its own voice and relays it to the resident.
- Arthur saves the answer to the property scoped knowledge base.
- Semantic matching means the next resident who asks the same thing in different words is answered straight away, interrupting nobody.
Managers view, edit, and delete every entry. That is the control point boards ask about. Residents get the answers management approved, and a changed rule is one edit. Escalation volume drops for a simple reason: each question a human answers is the last time a human answers it.
What makes a knowledge base property scoped?
It means the answers belong to one building and never leak across the portfolio. Your tower's guest parking rule does not become the default answer down the street. Every entry is tied to the property it was taught in.
It also differs from the uploaded FAQ file most chatbots call knowledge. Three differences matter:
- It grows from real conversations. Entries come from questions residents actually asked, not a document someone wrote at launch and never opened again.
- It sits beside live records, not instead of them. Arthur reads the documents, the booking calendar, the ticket queue, and the resident's own account. The knowledge base covers the judgment calls no record holds.
- It is editable. A wrong entry is deleted in a click. A changed rule is one edit. Nothing is buried inside a model you cannot inspect.
Documents work the same way. Bylaws, minutes, and budgets sit in the building's own document hub, indexed as they are added.
Why is acting different from answering?
Because an answer that does not become an action moves the work to tomorrow morning. A wrapper produces text. Arthur acts inside the system that runs the building.
Arthur creates service tickets, books amenities against the live calendar, registers parking, pre schedules visitors, RSVPs residents to events, sends forms, pulls receipts, sends payment links, and requests refunds that a property manager approves before money moves. Anything sensitive is PIN verified first. Arthur proposes and confirms, people approve, and nothing fires on its own.
This works because Arthur lives inside the platform holding the units, people, documents, vendors, and tickets. The context is native rather than fetched through an integration. Ask about the party room and Arthur reads the calendar the concierge uses. The ticket it opens is the ticket the manager sees. Answering and acting are one motion. We compared that design with a conversation layer bolted onto another system in EliseAI vs Arthur, and with an answering service. Arthur is the resident side only. HERO, URBI's operator facing agent, is the one managers use.
How does a generic AI wrapper compare with an orchestrator?
Not on intelligence. On where knowledge comes from, who controls it, and what survives the call.
| Dimension | Generic AI wrapper | Orchestrator with native context (Arthur in URBI) |
|---|---|---|
| Where the knowledge comes from | Public web text plus whatever FAQ file was uploaded at setup | The building's own records, documents, calendars, and a property scoped knowledge base |
| What happens when it does not know | Produces a plausible answer anyway, or dead ends the resident | Escalates to the property manager with a summary, instead of guessing |
| Who curates the answers | The model, and whoever last edited the prompt | Management, entry by entry, with view, edit, and delete on each |
| Does it improve | Same gap next week, same escalation | The approved answer is saved and semantically matched, so the next ask is handled alone |
| Can it act | No, it writes text and hands you a task | Tickets, bookings, parking, visitors, RSVPs, forms, receipts, payment links, refunds on approval |
| Who it will talk to about your unit | Whoever is typing | The verified caller, PIN gated, about their own household only |
| What record survives | A chat log, if you can export it | Transcript, summary, manager notification, and a permanent audit trail |
What safeguards should sit around the answers?
Controlling the answers is necessary but not sufficient. The rest of the risk sits in who is asking, where the conversation goes, and what is left behind.
- Household privacy scoping. Arthur answers about the caller's own household only. A neighbour cannot ask about your arrears or visitors.
- Consent first transfers. A transfer happens with the caller's agreement, and Arthur never dials a number the caller supplies.
- Sentiment scoring. Arthur reads the caller's mood in the background, shifts tone to match, and never cuts a caller off mid pause.
- A guaranteed record. Even on a mid call hangup, the transcript, summary, and manager notification are still produced. The 2am call is not lost.
- A permanent audit trail. Every action and AI tool call is logged, so when a board member asks how an answer was reached there is a record, not a memory.
None of this is exotic governance. The NIST AI Risk Management Framework, written to help organizations "better manage risks to individuals, organizations, and society associated with artificial intelligence", puts governance first and expects that "accountability structures are in place". A named human owning what the AI says is the point.
What should you ask any resident facing AI vendor?
Ask these seven in a live demo, not over email.
- Where does it get answers about my building, and can you show the source of one it just gave?
- What does it do when it does not know? Show that conversation live, not a slide.
- Can I see every answer it has learned for my property, then edit one and delete one while I watch?
- Does an answer taught at one property leak to another in the portfolio?
- Can it do anything beyond talk? Open a ticket, book the room, register the car.
- How does it verify the person asking is a resident, and what stops it discussing a neighbour's unit?
- What record exists after a call, including one that ends in the middle?
If a vendor cannot show numbers two, three, and five in a live system, you are buying a text generator with a property management logo on it. For a wider scan, see the best AI property management software, and URBI for residential buildings.
Frequently asked questions
Can we stop the AI from answering certain questions?
Yes. Anything the property has not taught, and no record covers, goes to the property manager instead of being answered. Managers curate the knowledge base directly, so an entry you do not want in circulation is deleted and stops being used. Legal interpretation, board deliberations, and anything a manager wants to own stay off the list.
Does the knowledge base replace our building documents?
No. Documents stay the source of truth for anything written down: bylaws, rules, minutes, budgets. The knowledge base covers practical questions no document answers cleanly, such as how your building handles a move in booked on short notice. Both feed Arthur, and every action lands in the same audit trail.
What happens on an emergency call at 2am?
Arthur answers immediately, opens the ticket in the real queue, routes it by category, and notifies the manager, instead of leaving a voicemail nobody hears until Monday. The transcript, summary, and notification exist even if the caller hangs up mid sentence. We covered the detail in whether AI handles emergency maintenance calls.
Any AI in front of your residents will speak with confidence. The only real question is whether management wrote what it says, and whether anything happens after it speaks. Arthur ships on URBI's Premium plan, month to month. To see the handoff, the knowledge base, and the audit trail in a live building, write to hello@myurbi.co.
